def build_v3(self):
"""
Normal Improved WGAN, with only real vs fake (no classes)
"""
self.z = tf.placeholder(tf.float32,
shape=[self.batch_size, self.data.latent_dim])
with tf.variable_scope("generator") as scope:
self.generator_output = Generator.generator(
self.z, self.is_training, self.gen_var_coll, self.gen_upd_coll)
with tf.variable_scope("discriminator") as scope:
disc_output_x = Discriminator.discriminator_v3(
self.x, self.batch_size, self.disc_var_coll)
scope.reuse_variables()
disc_output_gz = Discriminator.discriminator_v3(
self.generator_output, self.batch_size, self.disc_var_coll)
interpolates = tf.multiply(self.epsilon, self.x) + tf.multiply(
1 - self.epsilon, self.generator_output)
disc_interpolates = Discriminator.discriminator_v3(
interpolates, self.batch_size, self.disc_var_coll)
self.generator_loss = -tf.reduce_mean(disc_output_gz)
self.disc_loss = -tf.reduce_mean(disc_output_x) + tf.reduce_mean(
disc_output_gz)
gradients = tf.gradients(disc_interpolates, [interpolates])[0]
self.slopes = tf.sqrt(
tf.reduce_sum(tf.square(gradients), reduction_indices=[1]))
self.gradient_penalty = tf.reduce_mean((self.slopes - 1)**2)
self.disc_loss += self.lambdah * self.gradient_penalty
def build_v1(self):
"""
Version 1:
- D(x) has [num_classes+1] outputs
- D(x) uses one_hot class vector to compute its loss,
- G(z) uses one_hot class vector to generate, one_hot is the first [num_classes] elements of z
"""
self.z = tf.placeholder(
tf.float32, shape=[self.batch_size, self.data.latent_output_size])
with tf.variable_scope("generator") as scope:
self.generator_output = Generator.generator(
self.z, self.is_training, self.gen_var_coll, self.gen_upd_coll)
with tf.variable_scope("discriminator") as scope:
disc_output_x = Discriminator.discriminator_v1(
self.x, self.batch_size, self.num_classes, self.disc_var_coll)
scope.reuse_variables()
disc_output_gz = Discriminator.discriminator_v1(
self.generator_output, self.batch_size, self.num_classes,
disc_var_coll)
interpolates = tf.multiply(self.epsilon, self.x) + \
tf.multiply(1-self.epsilon, self.generator_output)
disc_interpolates = Discriminator.discriminator_v1(
interpolates, self.batch_size, self.num_classes, disc_var_coll)
# discriminator(self.generator_output_inner) will be of size:
# [batch_size, num_classes+1]
# labels will be of shape:
# [batch_size, num_classes+1]
self.generator_loss = tf.reduce_sum(tf.multiply(disc_output_gz, self.zlabels), axis=1) + \
tf.reduce_sum(tf.multiply(disc_output_gz, self.zlabels-1), axis=1)
batch_gen_loss = self.generator_loss
self.generator_loss = tf.reduce_mean(self.generator_loss)
self.disc_loss = tf.reduce_sum(tf.multiply(disc_output_x, self.xlabels), axis=1) + \
tf.reduce_sum(tf.multiply(disc_output_x, self.xlabels-1), axis=1) - batch_gen_loss
self.disc_loss = tf.reduce_mean(self.disc_loss)
gradients_per_dim = [
tf.gradients(
tf.slice(disc_interpolates, [0, i], [self.batch_size, 1]),
[interpolates]) for i in range(self.num_classes + 1)
]
slopes_per_dim = [
tf.sqrt(
tf.reduce_sum(tf.square(gradients_per_dim[i]),
reduction_indices=[1]))
for i in range(self.num_classes + 1)
]
gradient_penalty_per_dim = [
tf.reduce_mean((slopes_per_dim[i] - 1)**2)
for i in range(self.num_classes + 1)
]
total_grad_penalty = tf.zeros([])
for grad_penalty in gradient_penalty_per_dim:
total_grad_penalty += grad_penalty
self.disc_loss += self.lambdah * total_grad_penalty
def build_v2(self):
"""
Version 2:
- D(x) has 1 output
- D(x) takes in the one_hot class vector as an input to compute that 1 output
- G(z) takes in the one_hot class vector as before in Version 1
"""
self.z = tf.placeholder(
tf.float32, shape=[self.batch_size, self.data.latent_output_size])
with tf.variable_scope("generator") as scope:
self.generator_output = Generator.generator(
self.z, self.is_training, self.gen_var_coll, self.gen_upd_coll)
with tf.variable_scope("discriminator") as scope:
disc_output_x = Discriminator.discriminator_v2(
self.x, self.xlabels, self.batch_size, self.disc_var_coll, 50)
scope.reuse_variables()
disc_output_gz = Discriminator.discriminator_v2(
self.generator_output, self.zlabels, self.batch_size,
self.disc_var_coll, 50)
interpolates = tf.multiply(self.epsilon, self.x) + tf.multiply(
1 - self.epsilon, self.generator_output)
disc_interpolates = Discriminator.discriminator_v2(
interpolates, self.xlabels, self.batch_size,
self.disc_var_coll, 50)
self.generator_loss = tf.reduce_mean(disc_output_gz)
self.disc_loss = tf.reduce_mean(disc_output_x) - self.generator_loss
gradients = tf.gradients(disc_interpolates, [interpolates])[0]
slopes = tf.sqrt(
tf.reduce_sum(tf.square(gradients), reduction_indices=[1]))
gradient_penalty = tf.reduce_mean((slopes - 1)**2)
self.disc_loss += self.lambdah * gradient_penalty